I am building a deep learning model for NLP. I am pretty comfortable with adding word embedding from word2vec or Glove vectors as extra word features but I wanted to add other word features like POS tag of a word, NER tag of word along with embedding as features. How can I do this. Should I give these word features by concatenating their vector with the word vectors. Or is there some other method. Please suggest.
$\begingroup$
$\endgroup$
1
-
1$\begingroup$ Yes, concatenate them. $\endgroup$– EmreCommented Apr 28, 2017 at 21:52
Add a comment
|
2 Answers
$\begingroup$
$\endgroup$
0
One option is to concatenate them, the second is to treat them as separate inputs. For example Keras offers such neural model: https://keras.io/getting-started/functional-api-guide/#multi-input-and-multi-output-models
$\begingroup$
$\endgroup$
I would concatenate them into a single input vector. Essentially, your model treats each latent variable from the word embedding as a single feature (think about a regular ML model). Adding a couple to the end of this wouldn't hurt your performance too much.
Another option is to follow what @djstrong said, about multi-inputs. But I would start with just concatenating the extra variables at the end of your input vector.
